Conventional motion detection looks at a local error measure, commonly a sum-of-absolute-differences, between a target picture and a reference picture. Even if no motion exists, such local error measures tend to be non-zero due to noise and changes in scene lightness. Therefore, motion detection commonly detects small differences between the pictures as no motion and big differences as motion.
It would be desirable to implement level-based motion detection in a sequence of pictures.